NeuroNetFusion: enhanced EEG abnormality classification via multi-network TF-IDF feature selection.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalogram (EEG) signals present complex, nonlinear patterns challenging traditional classification methods.
• Existing EEG studies often rely on limited, uni-directional feature extraction, hindering comprehensive analysis.
• There is a need for advanced frameworks that capture multi-contextual dependencies within EEG data.

Purpose of the Study:

• To introduce NeuroNetFusion, a novel classification framework for EEG signals.
• To enhance bio-signal classification by integrating multi-directionally expressed cross-dependence information.
• To systematically integrate and select multi-context network-based features for improved EEG analysis.

Main Methods:

• EEG signal preprocessing using Savitzky-Golay (SG) filter and Discrete Wavelet Transform (DWT).
• Construction of causality networks (Directed Transfer Function - DTF) and correlation networks (Pearson correlation).
• Feature vectorization using TF-IDF and feature selection via a genetic algorithm.

Main Results:

• The NeuroNetFusion framework achieved a classification accuracy of 88.05% on the MTOUH dataset.
• Demonstrated an absolute improvement of 8.73% over the baseline Temporal Convolutional Network (TCN) model.
• Successfully bridged causality- and correlation-based representations for enhanced EEG interpretation.

Conclusions:

• NeuroNetFusion offers a significant advancement in EEG signal classification by leveraging integrated network features.
• The framework provides a scalable and interpretable pathway for analyzing cross-dependent bio-signal data.
• This approach holds promise for future applications in diagnosing neurological conditions and advancing bio-signal analysis.